Rumor Detection on Twitter with Tree-structured Recursive Neural Networks Jing Ma 1 , Wei Gao 2 , Kam-Fai Wong 1,3 1 The Chinese University of Hong Kong 2 Victoria University of Wellington, New Zealand 3 MoE Key Laboratory of High Confidence Software Technologies, China July 15-20, 2018 – ACL 2018 @ Melboume, Australia 1 Jing Ma (CUHK) 2018/7/15
Outline ￭ Introduction ￭ Related Work ￭ Problem Statement ￭ RvNN-based Rumor Detection ￭ Evaluation ￭ Conclusion and Future Work 2 Jing Ma (CUHK) 2018/7/15
Introduction What are rumors? A story or statement whose truth value is unverified or deliberately false 3 Jing Ma (CUHK) 2018/7/15
Introduction How the fake news propagated? Ø people tend to stop spreading a rumor if it is known as false. (Zubiaga et al., 2016b) Ø Previous studies focused on text mining from sequential microblog streams, we want to bridge the content semantics and denial propagation clues. supportive 4 Jing Ma (CUHK) 2016/7/1
Motivation ￭ We generally are not good at distinguishing rumors ￭ It is crucial to track and debunk rumors early to minimize their harmful effects. ￭ Online fact-checking services have limited topical coverage and long delay. ￭ Existing models use feature engineering – over simplistic; or recently deep neural networks – ignore propagation structures; Kernel-based method – develop based on tree structure but cannot learn high-level feature representations automatically. 5 Jing Ma (CUHK) 2018/7/15
Observation & Hypothesis ￭ Existing works: Consider post representation or propagation structure Support Neutral Doubt (a) RNN-based model (b) Tree kernel-based model (Ma et al. 2016) (Ma et al. 2017) ￭ IDEA: Combining the two models, leveraging propagation structure by representation learning algorithm 6 Jing Ma (CUHK) 2018/7/15
Observation & Hypothesis Why such model do better? Polarity stances (b) True rumor (a) False rumor Local characteristic: ￭ A reply usually respond to its immediate ancestor rather than the root tweet . ￭ Repliers tend to disagree with (or question) who support a false rumor or deny a true rumor; repliers tend to agree with who deny a false rumor or support a true rumor. 7 Jing Ma (CUHK) 2018/7/15
Contributions ￭ The first study that deeply integrates both structure and content semantics based on tree-structured recursive neural networks for detecting rumors from microblog posts ￭ Propose two variants of RvNN models based on bottom-up and top-down tree structures, to generate better integrated representations for a claim by capturing both structural and textural properties signaling rumors. ￭ Our experiments based on two real-world Twitter datasets achieve superior improvements over state-of-the-art baselines on both rumor classification and early detection tasks. ￭ We make the source codes in our experiments publicly accessible at https://github.com/majingCUHK/Rumor_RvNN 8 Jing Ma (CUHK) 2018/7/15
Outline ￭ Introduction ￭ Related Work ￭ Problem Statement ￭ RvNN-based Rumor Detection ￭ Evaluation ￭ Conclusion and Future Work 9 Jing Ma (CUHK) 2018/7/15
Related Work ￭ Systems based on common sense and investigative journalism, e.g., ￭ snopes.com ￭ factcheck.org ￭ Learning-based models for rumor detection ￭ Information credibility: Castillo et al. (2011), Yang et al. (2012) ￭ Using handcrafted and temporal features: Liu et al. (2015), Ma et al. (2015), Kwon et al. (2013, 2017) ￭ Using cue terms: Zhao et al. (2015) ￭ Using recurrent neural networks: Ma et al. (2016, 2018) Without hand- ￭ Tree-kernel-based model: crafted features Ma et al. (2017), Wu et al. (2015) ￭ RvNN-based works ￭ images segmentation (Socher et al, 2011) ￭ phrase representation from word vectors (Socher et al, 2012) ￭ Sentiment analysis (Socher et al, 2013) ￭ etc 10 Jing Ma (CUHK) 2018/7/15
Outline ￭ Introduction ￭ Related Work ￭ Problem Statement ￭ RvNN-based Rumor Detection ￭ Evaluation ￭ Conclusion and Future Work 11 Jing Ma (CUHK) 2018/7/15
Problem Statement ￭ Given a set of microblog posts R = {𝑠} , model each source tweet as a tree structure T 𝑠 = < 𝑊, 𝐹 > , where each node 𝑤 provide the text content of each post. And 𝐹 is directed edges corresponding to response relation. ￭ Task 1 – finer-grained classification for each source post false rumor, true rumor, non-rumor, unverified rumor ￭ Task 2 – detect rumor as early as possible 12 Jing Ma (CUHK) 2018/7/15
Tweet Structure Root tweet 𝒚 𝟐 : #Walmart donates $10,000 to #DarrenWilson bottom-up tree fund to continue police racial profiling… 𝒚 𝟑 : 1:30 Idc if they killed 𝒚 𝟒 : NEED SOURCE. a mf foreal. Ima always have a feeling this shop with @Walmart. I'm is just hearsay ... just bein honest 💂 𝒚 𝟓 : I agree. I have been 𝒚 𝟔 : Exactly, i don't think hearing this all day but Wal-Mart would let everyone no source 1:12 know this if they did!! 2:21 replies 𝒚 𝟐 : #Walmart donates $10,000 to #DarrenWilson top-down tree fund to continue police racial profiling… 𝒚 𝟑 : 1:30 Idc if they killed 𝒚 𝟒 : NEED SOURCE. a mf foreal. Ima always have a feeling this shop with @Walmart. I'm is just hearsay ... just bein honest 💂 𝒚 𝟓 : I agree. I have been 𝒚 𝟔 : Exactly, i don't think hearing this all day but Wal-Mart would let everyone no source 1:12 know this if they did!! 2:21 13 2018/7/15 Jing Ma (CUHK)
Outline ￭ Introduction ￭ Related Work ￭ Problem Statement ￭ RvNN-based Rumor Detection ￭ Evaluation ￭ Conclusion and Future Work 14 Jing Ma (CUHK) 2018/7/15
Standard Recursive Neural Networks ￭ RvNN ( tree-structured neural networks ) utilize sentence parse trees: representation associated with each node of a parse tree is computed from its direct children, computed by 𝑞 = 𝑔(𝑋 9 𝑑 ; ; 𝑑 = + 𝑐) ￭ p : the feature vector of a parent node whose children are 𝑑 ; and 𝑑 = ￭ computation is done recursively over all tree nodes 15 Jing Ma (CUHK) 2018/7/15
Bottom-up RvNN Ø Input: bottom-up tree (node: a post represented as a vector of words ) Ø GRU equation at node 𝑘 Children node Ø Structure : recursively visit every node from the leaves at the bottom to the root at the top (a natural extension to the original RvNN ) Own input Ø Intuition : local rumor indicative features are aggregated along different branches (e.g., subtrees having a denial parent and a set of supportive children) (generate a feature vector for each subtree) 𝒚 𝟐 : #Walmart donates $10,000 to #DarrenWilson fund to continue police racial profiling… 𝒚 𝟑 : 1:30 Idc if they killed 𝒚 𝟒 : NEED SOURCE. a mf foreal. Ima always have a feeling this shop with @Walmart. I'm is just hearsay ... just bein honest 💂 𝒚 𝟓 : I agree. I have been 𝒚 𝟔 : Exactly, i don't think hearing this all day but Wal-Mart would let everyone no source 1:12 know this if they did!! 2:21 16 Jing Ma (CUHK) 2018/7/15
Top-down RvNN Ø GRU transition equation at node 𝑘 Ø Input: top-down tree Own input Parent node Ø Structure : recursively visit from the root node to its children until reaching all leaf nodes. (reverse Bottom-up RvNN) Ø Intuition : rumor-indicative features are aggregated along the propagation path (e.g., if a post agree with its parent’s stance, the parent’s stance should be reinforced) (models how information flows from source post to the current node) 𝒚 𝟐 : #Walmart donates $10,000 to #DarrenWilson fund to continue police racial profiling… 𝒚 𝟑 : 1:30 Idc if they killed 𝒚 𝟒 : NEED SOURCE. a mf foreal. Ima always have a feeling this shop with @Walmart. I'm is just hearsay ... just bein honest 💂 𝒚 𝟓 : I agree. I have been 𝒚 𝟔 : Exactly, i don't think hearing this all day but Wal-Mart would let everyone no source 1:12 know this if they did!! 2:21 17 Jing Ma (CUHK) 2018/7/15
Model Training Ø Comparison: both of the two RvNN models aim to capture the structural properties by recursively visiting all nodes Bottom-up RvNN : the state of root node (i.e., source tweet) can be regard as the representation of the whole tree (can be used for supervised classification). Top-down RvNN : the representation of each path are eventually embedded into the hidden vector of all the leaf nodes. learned vector of root node Ø Output Layer Bottom-up RvNN: 𝑧 = 𝑇𝑝𝑔𝑢𝑛𝑏𝑦 𝑊ℎ K + 𝑐 Top-down RvNN: 𝑧 = 𝑇𝑝𝑔𝑢𝑛𝑏𝑦 𝑊ℎ L + 𝑐 the pooling vector over all leaf nodes T R Q O = = Ø Objective Function: 𝑀 = ∑ ∑ 𝑧 O − 𝑧 + 𝜇 Θ = US; OS; prediction Ground truth Ø Training Procedure parameters are updated using efficient back-propagation through structure ( Goller and Kuchler, 1996; Socher et al., 2013 ) 18 Jing Ma (CUHK) 2018/7/15
Outline ￭ Introduction ￭ Related Work ￭ Problem Statement ￭ RvNN-based Rumor Detection ￭ Evaluation ￭ Conclusion and Future Work 19 Jing Ma (CUHK) 2018/7/15
Data Collection ￭ Use two reference Tree datasets: URL of the datasets: https://www.dropbox.com/s/0jhsfwep3ywvpca/rumdetect2017.zip?dl=0 20 Jing Ma (CUHK) 2018/7/15
Recommend
More recommend
